Prime Radiant

evener tags v0.3.8 after a week-long copy-oracle test-hardening push, and evener-client-playdate debuts as a native Playdate handheld client

Week of October 12, 2026 · Oct 12, 2026–Oct 18, 2026 · published Oct 16, 2026

prime-radiant-inc/evener tags v0.3.8 this week, after a week running a sustained “copy-oracle” test-hardening campaign alongside a web-UI widget-library migration, device-pairing work, and an incrementally hardened per-test timing gate. A new repo, prime-radiant-inc/evener-client-playdate, also debuts this week as a native Playdate handheld client for evener, with voice composition and direct Tailscale connectivity. This week is still open: the data below runs through Friday evening, not the full calendar week.

The week in numbers

  • 7 repos active (2 featured, 5 also shipped)
  • 236 commits
  • 207 merged PRs
  • +188,178 / −56,675 lines
  • 3 commit identities
  • 1 tagged release

prime-radiant-inc/evener

A coding agent: give it a prompt and it reads, writes, runs commands, and searches code in a loop until the work is done, using native tool-calling across OpenAI, Anthropic, and Google models.

191 commits 190 PRs 🏷 v0.3.8 +179,552 −51,413

Tagged release v0.3.8. By far the most active repo this week. The work clusters into a few sustained threads rather than one single feature: a large “copy-oracle” test-hardening campaign spanning nearly the whole week and touching web, native, hub, agent, TUI, sandbox, frontend, and appwire-client tests — converting assertions on literal copy/wording into assertions on structural state, markers, error sentinels/types, or toast kinds, backed by new tooling to catch regressions of the same kind (a copy-oracle inventory of wording-only tests, a CI ratchet against new ones, and that inventory tool itself later ported from Python to Go and TypeScript); a web-UI migration onto the shared widget library, done in explicit waves with a final follow-up and “last leftovers” pass; device/mobile pairing work (shared device invitations plus native Hub pairing, pairing destinations kept visible on narrow screens, automatic Hub-destination discovery for mobile pairing, clarified phone QR/sound-pairing codes); a per-test timing “gate” hardened incrementally to fail loudly on a malformed testing-budget config instead of silently, then segmented by test class, with some runner-bound classes returned to warn-only per a dated ruling; and entry-backed background transcripts, landed as a feature and then followed by a repair pass for defects the review found. Alongside those: memory work (pages that say when to recall themselves, a memory_write taking when/body fields, index lines linking the page path instead of the title), execenv/sandbox read-capping, several LLM/vision provider fixes, hub reliability fixes, and a run of changes landing right at the window’s close: Stop taking the composer’s slot with the status line lifting to the title, a floating compose button on the Board replacing its bottom bar, the session UI trimmed (dropping the like-this control for a copyable id and a smaller composer), settings reorganized with plugin sources unified, and delegate report messages preserved in more detail. Sole author: Jesse Vincent (staff, obra).

prime-radiant-inc/evener-client-playdate

Native Playdate client for Evener with voice composition and direct Tailscale connectivity.

18 commits 14 PRs

A new repo, created this week — a native client for the Playdate handheld, built out through a series of same-day-merged PRs: an initial qualified import of the client made independent of the main evener checkout, then a run of pairing work (a shared full device-pairing fingerprint, documented automatic Hub destinations, clarified sound-pairing comparison codes), investigating and fixing Playdate TCP stalls with work preserved across runtime recovery, reduced read/send delays and idle redraws, conversation content loaded through history-inclusion flags, longer conversations made readable sooner with adjustable text, microphone-input display during setup-sound pairing, acoustic setup decoded across symbol-timing offsets, and — toward the end of the window — a recorded onboarding walkthrough, promo footage rendered from captured clips, the Playdate crank rendered in its correct side-wall plane, and both films embedded as inline README video players. No releases yet. Sole author: Jesse Vincent (staff, obra).

Also shipped

prime-radiant-inc/maintaining-documentation

A coding-agent skill that keeps a project’s docs true to the code and useful to their readers, for every harness.

4 commits 1 PR +2,477 −489

A young repo (created just before this window opened), with one merged PR this window — reader-centered doc writing: a reader brief, doc types, and a reader test. Also: a copyedit pass that runs a fresh subagent against the project’s bundled Strunk guide and then a second pass checking the resulting word-diff for changed claims, evaluated in blind reader comparisons; the writing-clearly-and-concisely skill bundled in wholesale from obra/the-elements-of-style (public-domain Strunk text); then renamed to copyediting-docs to avoid colliding with the standalone skill, with its guide trimmed — composition rules kept word-for-word, punctuation/usage trimmed by deletion only, reaching parity against full-text copyedits in blind judging. No tags or releases on this repo. Sole author: Jesse Vincent (staff, obra).

obra/lace

Lightweight agentic coding environment.

2 commits 2 PRs +854 −75

A small week: a fix so a prompt retried after the agent process died mid-turn gets a new turn run instead of being treated as already-handled, and namespaced skill directories — SkillRegistry now accepts { dir, namespace } entries alongside bare strings, every parsed skill version is kept under its qualified (namespace:name) form, plugin skill dirs keep their plugin namespace while persona/bundled-core dirs stay un-namespaced, and <available_skills>/use_skill resolve either bare or qualified names. Sole contributor: Jesse Vincent (staff, obra).

obra/picksandshovels

PICKS & SHOVELS: a pixel-art game about the AI gold rush. Sell the GPUs, keep p(doom) under 100%.

7 commits +575 −101

A personal project, not a company product, in its first real week of commits (committed directly, no PRs, no releases). Work: sharing (one-click copy of image + linked caption, separated image/text copy, a visible caption, native share sheet) and analytics — production-only Google Analytics with game_start/game_end/share events, fuller journey analytics (screen views, run milestones, yearly checkpoints, choices, bosses, quits/abandons, user properties) moved onto a dedicated GA property, and a fix for a page-view being double-counted between the tag’s automatic page_view and the journey’s own named screen-view. Sole author: Jesse Vincent (staff, obra).

obra/muse-skills

No repository description set.

4 commits

Commits look automated (a daily cadence, under a git identity with no linked GitHub account); not confirmed. No human commits to attribute.

The remaining active repo — claude-plugin-stats — had only bot-driven activity again this week (five days of the daily install-stats scrape and chart rebuild), with no human commits.